{
  "domain": "ai.onnx",
  "name": "BitCast",
  "sinceVersion": 26,
  "inputs": { "input": { "dtype": "T" } },
  "outputs": { "output": { "dtype": "U", "rank": "ranks.input", "shape": "shapes.input" } },
  "attributes": { "to": {} },
  "attributeConstraints": { "to": { "required": true } },
  "typeConstraints": {
    "T": ["float32", "int8", "int32", "uint8", "uint32"],
    "U": ["float32", "int8", "int32", "uint8", "uint32"]
  },
  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
  "derive": {
    "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
    "bitcastShapeOk": "ranks.input >= 0 and ranks.output == ranks.input and numel(shapes.input) == numel(shapes.output)",
    "bitcastTypeOk": "attrs.to == onnxDtypeCode(logicalDtypes.U) and (((tensorDtypes.input == \"int8\" or tensorDtypes.input == \"uint8\") and (tensorDtypes.output == \"int8\" or tensorDtypes.output == \"uint8\")) or (tensorDtypes.input != \"int8\" and tensorDtypes.input != \"uint8\" and tensorDtypes.output != \"int8\" and tensorDtypes.output != \"uint8\"))",
    "bitcastWorkgroupSize": "min(tunables.WORKGROUP_SIZE, deviceWorkgroupCap)",
    "bitcastDispatchFits": "ceilDiv(ceilDiv(numel(shapes.output), bitcastWorkgroupSize), min(device.limits.maxComputeWorkgroupsPerDimension, 65535)) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)",
    "bitcastBaseOk": "bitcastShapeOk and bitcastTypeOk and bitcastDispatchFits",
    "inScalar": "dtypes.T",
    "outScalar": "dtypes.U",
    "inputIsInt8": "tensorDtypes.input == \"int8\"",
    "inputIsUint8": "tensorDtypes.input == \"uint8\"",
    "outputIsInt8": "tensorDtypes.output == \"int8\"",
    "outputIsUint8": "tensorDtypes.output == \"uint8\""
  },
  "when": ["bitcastBaseOk"],
  "variants": [
    {
      "id": "slot32_vec4",
      "priority": 20,
      "when": ["numel(shapes.input) % 4 == 0"],
      "derive": { "inVec4": "\"vec4<\" ~ dtypes.T ~ \">\"", "outVec4": "\"vec4<\" ~ dtypes.U ~ \">\"" },
      "passes": [
        {
          "id": "main",
          "name": "BitCast.vec4",
          "shader": "bitcast.wgsl.jinja",
          "derive": { "vectorizedSpec": true, "workgroupSizeSpec": "bitcastWorkgroupSize" },
          "bindings": [
            { "arg": "input", "elementType": "$inVec4" },
            { "arg": "output", "elementType": "$outVec4" },
            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.output) / 4" }] }
          ],
          "dispatch": {
            "x": "min(ceilDiv((numel(shapes.output) / 4), (bitcastWorkgroupSize)), 65535)",
            "y": "ceilDiv(ceilDiv((numel(shapes.output) / 4), (bitcastWorkgroupSize)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "slot32",
      "passes": [
        {
          "id": "main",
          "name": "BitCast",
          "shader": "bitcast.wgsl.jinja",
          "derive": { "workgroupSizeSpec": "bitcastWorkgroupSize" },
          "bindings": [
            "input",
            "output",
            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.output)" }] }
          ],
          "dispatch": {
            "x": "min(ceilDiv((numel(shapes.output)), (bitcastWorkgroupSize)), 65535)",
            "y": "ceilDiv(ceilDiv((numel(shapes.output)), (bitcastWorkgroupSize)), 65535)",
            "z": 1
          }
        }
      ]
    }
  ]
}